The terminal graph of matrix multiplication: definitions #
The 3 n² terminals of n × n matrix multiplication C = A B, the inputs A i j and B j k
and the outputs C i k, are the edges of the complete tripartite graph on three copies I,
J, K of Fin n: A i j joins i ∈ I to j ∈ J, B j k joins j ∈ J to k ∈ K, and
C i k joins i ∈ I to k ∈ K. Every vertex has degree 2 n.
A split of the wires of a circuit places some terminals on its side S. We record the placed
terminals of each kind as a set of index pairs: a holds the pairs (i, j) with A i j placed,
b the pairs (j, k) with B j k placed, and c the pairs (i, k) with C i k placed.
rowSet a iandcolSet a j: the second coordinates paired withi, and the first coordinates paired withj.degI a c i,degJ a b j,degK b c k: the number of placed terminals at a vertex.minority n d = min d (2 n - d): the smaller of the numbers of terminals at a vertex of degree2 non the two sides, whendof them are placed.chargeI a c,chargeJ a b,chargeK b c: the sums ofminorityover the vertices ofI,JandK. For matrix multiplication they are lower bounds for ranks of three linear maps read off the split: fixingB(chargeI), fixingA(chargeK), and the cross block of the Hessian of∑ Λ i k C i k(chargeJ).heavyI a c,heavyJ a b,heavyK b c: the heavy vertices, with at leastnplaced terminals;heavyCount a b ccounts them.excessI a c i,deficitI a c i: the excessdeg - nand deficitn - degof a vertex ofIrelative to the thresholdn.colSetIn a R jandcomplColSetIn a R j: the rowsi ∈ RwithA i jplaced, resp. unplaced.probeIJ a b c I₀andtwoSubsetProbeIJ a b c I₀ I₁: exploratory sums mixing theIandJcharges along subsets ofI. They are combinatorial quantities only; no rank-cut theorem relates them to the signals crossing a split.mixedPairs X Y: the pairs(x, y)with exactly one ofx ∈ Xandy ∈ Y.lightHeavy a b c: the number of terminals with exactly one heavy endpoint.place t S: the index pairs whose terminal wiret plies in a set of wiresS.
The smaller side at a vertex of degree 2 n with d placed terminals.
Equations
- Algebraic.Cutwidth.MultiOutput.Tripartite.minority n d = min d (2 * n - d)
Instances For
The charge of the vertices of I.
Equations
Instances For
The charge of the vertices of J.
Equations
Instances For
The charge of the vertices of K.
Equations
Instances For
The heavy vertices of I: at least n placed terminals.
Equations
- Algebraic.Cutwidth.MultiOutput.Tripartite.heavyI a c = {i : Fin n | n ≤ Algebraic.Cutwidth.MultiOutput.Tripartite.degI a c i}
Instances For
The heavy vertices of J.
Equations
- Algebraic.Cutwidth.MultiOutput.Tripartite.heavyJ a b = {j : Fin n | n ≤ Algebraic.Cutwidth.MultiOutput.Tripartite.degJ a b j}
Instances For
The heavy vertices of K.
Equations
- Algebraic.Cutwidth.MultiOutput.Tripartite.heavyK b c = {k : Fin n | n ≤ Algebraic.Cutwidth.MultiOutput.Tripartite.degK b c k}
Instances For
The first coordinates in R paired with j in a: the placed terminals A i j with
i ∈ R.
Equations
- Algebraic.Cutwidth.MultiOutput.Tripartite.colSetIn a R j = {x ∈ Algebraic.Cutwidth.MultiOutput.Tripartite.colSet a j | x ∈ R}
Instances For
The first coordinates in R not paired with j in a: the unplaced terminals A i j with
i ∈ R.
Equations
- Algebraic.Cutwidth.MultiOutput.Tripartite.complColSetIn a R j = {x ∈ R | x ∉ Algebraic.Cutwidth.MultiOutput.Tripartite.colSet a j}
Instances For
An exploratory probe mixing the I and J charges along a subset I₀ of I: the
I-charge minority n (degI a c i) of the rows i ∈ I₀, plus, at each j ∈ J, the two summands
of minority n (degJ a b j) = min x (n - y) + min (n - x) y (minority_add_eq, with
x = (colSet a j).card, y = (rowSet b j).card) with the A-terminals restricted to the rows of
I₀ᶜ. It interpolates between chargeJ (I₀ = ∅) and chargeI (I₀ = univ).
This is a combinatorial quantity only: no theorem bounds the number of signals crossing a split below by it.
Equations
- One or more equations did not get rendered due to their size.
Instances For
A two-subset variant of probeIJ. At i ∈ I, minority n (degI a c i) is the sum of
min r (n - s) and min (n - r) s (minority_add_eq, with r = (rowSet a i).card,
s = (rowSet c i).card); here the first summand is counted over I₀ and the second over I₁.
At each j ∈ J, the two summands of minority n (degJ a b j) are counted with the
A-terminals restricted to the rows of I₀ᶜ and of I₁ᶜ respectively. As for probeIJ, the
arguments are the rows charged on the I side, and
twoSubsetProbeIJ a b c I₀ I₀ = probeIJ a b c I₀.
This is a combinatorial quantity only: no theorem bounds the number of signals crossing a split below by it.
Equations
- One or more equations did not get rendered due to their size.